SUNRPC: Convert the xprt->sending queue back to an ordinary wait queue
[linux-2.6-block.git] / net / sunrpc / xprt.c
1 /*
2  *  linux/net/sunrpc/xprt.c
3  *
4  *  This is a generic RPC call interface supporting congestion avoidance,
5  *  and asynchronous calls.
6  *
7  *  The interface works like this:
8  *
9  *  -   When a process places a call, it allocates a request slot if
10  *      one is available. Otherwise, it sleeps on the backlog queue
11  *      (xprt_reserve).
12  *  -   Next, the caller puts together the RPC message, stuffs it into
13  *      the request struct, and calls xprt_transmit().
14  *  -   xprt_transmit sends the message and installs the caller on the
15  *      transport's wait list. At the same time, if a reply is expected,
16  *      it installs a timer that is run after the packet's timeout has
17  *      expired.
18  *  -   When a packet arrives, the data_ready handler walks the list of
19  *      pending requests for that transport. If a matching XID is found, the
20  *      caller is woken up, and the timer removed.
21  *  -   When no reply arrives within the timeout interval, the timer is
22  *      fired by the kernel and runs xprt_timer(). It either adjusts the
23  *      timeout values (minor timeout) or wakes up the caller with a status
24  *      of -ETIMEDOUT.
25  *  -   When the caller receives a notification from RPC that a reply arrived,
26  *      it should release the RPC slot, and process the reply.
27  *      If the call timed out, it may choose to retry the operation by
28  *      adjusting the initial timeout value, and simply calling rpc_call
29  *      again.
30  *
31  *  Support for async RPC is done through a set of RPC-specific scheduling
32  *  primitives that `transparently' work for processes as well as async
33  *  tasks that rely on callbacks.
34  *
35  *  Copyright (C) 1995-1997, Olaf Kirch <okir@monad.swb.de>
36  *
37  *  Transport switch API copyright (C) 2005, Chuck Lever <cel@netapp.com>
38  */
39
40 #include <linux/module.h>
41
42 #include <linux/types.h>
43 #include <linux/interrupt.h>
44 #include <linux/workqueue.h>
45 #include <linux/net.h>
46 #include <linux/ktime.h>
47
48 #include <linux/sunrpc/clnt.h>
49 #include <linux/sunrpc/metrics.h>
50 #include <linux/sunrpc/bc_xprt.h>
51 #include <linux/rcupdate.h>
52
53 #include <trace/events/sunrpc.h>
54
55 #include "sunrpc.h"
56
57 /*
58  * Local variables
59  */
60
61 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
62 # define RPCDBG_FACILITY        RPCDBG_XPRT
63 #endif
64
65 /*
66  * Local functions
67  */
68 static void      xprt_init(struct rpc_xprt *xprt, struct net *net);
69 static __be32   xprt_alloc_xid(struct rpc_xprt *xprt);
70 static void     xprt_connect_status(struct rpc_task *task);
71 static void      xprt_destroy(struct rpc_xprt *xprt);
72
73 static DEFINE_SPINLOCK(xprt_list_lock);
74 static LIST_HEAD(xprt_list);
75
76 /**
77  * xprt_register_transport - register a transport implementation
78  * @transport: transport to register
79  *
80  * If a transport implementation is loaded as a kernel module, it can
81  * call this interface to make itself known to the RPC client.
82  *
83  * Returns:
84  * 0:           transport successfully registered
85  * -EEXIST:     transport already registered
86  * -EINVAL:     transport module being unloaded
87  */
88 int xprt_register_transport(struct xprt_class *transport)
89 {
90         struct xprt_class *t;
91         int result;
92
93         result = -EEXIST;
94         spin_lock(&xprt_list_lock);
95         list_for_each_entry(t, &xprt_list, list) {
96                 /* don't register the same transport class twice */
97                 if (t->ident == transport->ident)
98                         goto out;
99         }
100
101         list_add_tail(&transport->list, &xprt_list);
102         printk(KERN_INFO "RPC: Registered %s transport module.\n",
103                transport->name);
104         result = 0;
105
106 out:
107         spin_unlock(&xprt_list_lock);
108         return result;
109 }
110 EXPORT_SYMBOL_GPL(xprt_register_transport);
111
112 /**
113  * xprt_unregister_transport - unregister a transport implementation
114  * @transport: transport to unregister
115  *
116  * Returns:
117  * 0:           transport successfully unregistered
118  * -ENOENT:     transport never registered
119  */
120 int xprt_unregister_transport(struct xprt_class *transport)
121 {
122         struct xprt_class *t;
123         int result;
124
125         result = 0;
126         spin_lock(&xprt_list_lock);
127         list_for_each_entry(t, &xprt_list, list) {
128                 if (t == transport) {
129                         printk(KERN_INFO
130                                 "RPC: Unregistered %s transport module.\n",
131                                 transport->name);
132                         list_del_init(&transport->list);
133                         goto out;
134                 }
135         }
136         result = -ENOENT;
137
138 out:
139         spin_unlock(&xprt_list_lock);
140         return result;
141 }
142 EXPORT_SYMBOL_GPL(xprt_unregister_transport);
143
144 /**
145  * xprt_load_transport - load a transport implementation
146  * @transport_name: transport to load
147  *
148  * Returns:
149  * 0:           transport successfully loaded
150  * -ENOENT:     transport module not available
151  */
152 int xprt_load_transport(const char *transport_name)
153 {
154         struct xprt_class *t;
155         int result;
156
157         result = 0;
158         spin_lock(&xprt_list_lock);
159         list_for_each_entry(t, &xprt_list, list) {
160                 if (strcmp(t->name, transport_name) == 0) {
161                         spin_unlock(&xprt_list_lock);
162                         goto out;
163                 }
164         }
165         spin_unlock(&xprt_list_lock);
166         result = request_module("xprt%s", transport_name);
167 out:
168         return result;
169 }
170 EXPORT_SYMBOL_GPL(xprt_load_transport);
171
172 static void xprt_clear_locked(struct rpc_xprt *xprt)
173 {
174         xprt->snd_task = NULL;
175         if (!test_bit(XPRT_CLOSE_WAIT, &xprt->state)) {
176                 smp_mb__before_atomic();
177                 clear_bit(XPRT_LOCKED, &xprt->state);
178                 smp_mb__after_atomic();
179         } else
180                 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
181 }
182
183 /**
184  * xprt_reserve_xprt - serialize write access to transports
185  * @task: task that is requesting access to the transport
186  * @xprt: pointer to the target transport
187  *
188  * This prevents mixing the payload of separate requests, and prevents
189  * transport connects from colliding with writes.  No congestion control
190  * is provided.
191  */
192 int xprt_reserve_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
193 {
194         struct rpc_rqst *req = task->tk_rqstp;
195
196         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
197                 if (task == xprt->snd_task)
198                         return 1;
199                 goto out_sleep;
200         }
201         if (test_bit(XPRT_WRITE_SPACE, &xprt->state))
202                 goto out_unlock;
203         xprt->snd_task = task;
204
205         return 1;
206
207 out_unlock:
208         xprt_clear_locked(xprt);
209 out_sleep:
210         dprintk("RPC: %5u failed to lock transport %p\n",
211                         task->tk_pid, xprt);
212         task->tk_timeout = RPC_IS_SOFT(task) ? req->rq_timeout : 0;
213         task->tk_status = -EAGAIN;
214         rpc_sleep_on(&xprt->sending, task, NULL);
215         return 0;
216 }
217 EXPORT_SYMBOL_GPL(xprt_reserve_xprt);
218
219 static bool
220 xprt_need_congestion_window_wait(struct rpc_xprt *xprt)
221 {
222         return test_bit(XPRT_CWND_WAIT, &xprt->state);
223 }
224
225 static void
226 xprt_set_congestion_window_wait(struct rpc_xprt *xprt)
227 {
228         if (!list_empty(&xprt->xmit_queue)) {
229                 /* Peek at head of queue to see if it can make progress */
230                 if (list_first_entry(&xprt->xmit_queue, struct rpc_rqst,
231                                         rq_xmit)->rq_cong)
232                         return;
233         }
234         set_bit(XPRT_CWND_WAIT, &xprt->state);
235 }
236
237 static void
238 xprt_test_and_clear_congestion_window_wait(struct rpc_xprt *xprt)
239 {
240         if (!RPCXPRT_CONGESTED(xprt))
241                 clear_bit(XPRT_CWND_WAIT, &xprt->state);
242 }
243
244 /*
245  * xprt_reserve_xprt_cong - serialize write access to transports
246  * @task: task that is requesting access to the transport
247  *
248  * Same as xprt_reserve_xprt, but Van Jacobson congestion control is
249  * integrated into the decision of whether a request is allowed to be
250  * woken up and given access to the transport.
251  * Note that the lock is only granted if we know there are free slots.
252  */
253 int xprt_reserve_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
254 {
255         struct rpc_rqst *req = task->tk_rqstp;
256
257         if (test_and_set_bit(XPRT_LOCKED, &xprt->state)) {
258                 if (task == xprt->snd_task)
259                         return 1;
260                 goto out_sleep;
261         }
262         if (req == NULL) {
263                 xprt->snd_task = task;
264                 return 1;
265         }
266         if (test_bit(XPRT_WRITE_SPACE, &xprt->state))
267                 goto out_unlock;
268         if (!xprt_need_congestion_window_wait(xprt)) {
269                 xprt->snd_task = task;
270                 return 1;
271         }
272 out_unlock:
273         xprt_clear_locked(xprt);
274 out_sleep:
275         dprintk("RPC: %5u failed to lock transport %p\n", task->tk_pid, xprt);
276         task->tk_timeout = RPC_IS_SOFT(task) ? req->rq_timeout : 0;
277         task->tk_status = -EAGAIN;
278         rpc_sleep_on(&xprt->sending, task, NULL);
279         return 0;
280 }
281 EXPORT_SYMBOL_GPL(xprt_reserve_xprt_cong);
282
283 static inline int xprt_lock_write(struct rpc_xprt *xprt, struct rpc_task *task)
284 {
285         int retval;
286
287         if (test_bit(XPRT_LOCKED, &xprt->state) && xprt->snd_task == task)
288                 return 1;
289         spin_lock_bh(&xprt->transport_lock);
290         retval = xprt->ops->reserve_xprt(xprt, task);
291         spin_unlock_bh(&xprt->transport_lock);
292         return retval;
293 }
294
295 static bool __xprt_lock_write_func(struct rpc_task *task, void *data)
296 {
297         struct rpc_xprt *xprt = data;
298
299         xprt->snd_task = task;
300         return true;
301 }
302
303 static void __xprt_lock_write_next(struct rpc_xprt *xprt)
304 {
305         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
306                 return;
307         if (test_bit(XPRT_WRITE_SPACE, &xprt->state))
308                 goto out_unlock;
309         if (rpc_wake_up_first_on_wq(xprtiod_workqueue, &xprt->sending,
310                                 __xprt_lock_write_func, xprt))
311                 return;
312 out_unlock:
313         xprt_clear_locked(xprt);
314 }
315
316 static void __xprt_lock_write_next_cong(struct rpc_xprt *xprt)
317 {
318         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
319                 return;
320         if (test_bit(XPRT_WRITE_SPACE, &xprt->state))
321                 goto out_unlock;
322         if (xprt_need_congestion_window_wait(xprt))
323                 goto out_unlock;
324         if (rpc_wake_up_first_on_wq(xprtiod_workqueue, &xprt->sending,
325                                 __xprt_lock_write_func, xprt))
326                 return;
327 out_unlock:
328         xprt_clear_locked(xprt);
329 }
330
331 /**
332  * xprt_release_xprt - allow other requests to use a transport
333  * @xprt: transport with other tasks potentially waiting
334  * @task: task that is releasing access to the transport
335  *
336  * Note that "task" can be NULL.  No congestion control is provided.
337  */
338 void xprt_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
339 {
340         if (xprt->snd_task == task) {
341                 xprt_clear_locked(xprt);
342                 __xprt_lock_write_next(xprt);
343         }
344 }
345 EXPORT_SYMBOL_GPL(xprt_release_xprt);
346
347 /**
348  * xprt_release_xprt_cong - allow other requests to use a transport
349  * @xprt: transport with other tasks potentially waiting
350  * @task: task that is releasing access to the transport
351  *
352  * Note that "task" can be NULL.  Another task is awoken to use the
353  * transport if the transport's congestion window allows it.
354  */
355 void xprt_release_xprt_cong(struct rpc_xprt *xprt, struct rpc_task *task)
356 {
357         if (xprt->snd_task == task) {
358                 xprt_clear_locked(xprt);
359                 __xprt_lock_write_next_cong(xprt);
360         }
361 }
362 EXPORT_SYMBOL_GPL(xprt_release_xprt_cong);
363
364 static inline void xprt_release_write(struct rpc_xprt *xprt, struct rpc_task *task)
365 {
366         if (xprt->snd_task != task)
367                 return;
368         spin_lock_bh(&xprt->transport_lock);
369         xprt->ops->release_xprt(xprt, task);
370         spin_unlock_bh(&xprt->transport_lock);
371 }
372
373 /*
374  * Van Jacobson congestion avoidance. Check if the congestion window
375  * overflowed. Put the task to sleep if this is the case.
376  */
377 static int
378 __xprt_get_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
379 {
380         if (req->rq_cong)
381                 return 1;
382         dprintk("RPC: %5u xprt_cwnd_limited cong = %lu cwnd = %lu\n",
383                         req->rq_task->tk_pid, xprt->cong, xprt->cwnd);
384         if (RPCXPRT_CONGESTED(xprt)) {
385                 xprt_set_congestion_window_wait(xprt);
386                 return 0;
387         }
388         req->rq_cong = 1;
389         xprt->cong += RPC_CWNDSCALE;
390         return 1;
391 }
392
393 /*
394  * Adjust the congestion window, and wake up the next task
395  * that has been sleeping due to congestion
396  */
397 static void
398 __xprt_put_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
399 {
400         if (!req->rq_cong)
401                 return;
402         req->rq_cong = 0;
403         xprt->cong -= RPC_CWNDSCALE;
404         xprt_test_and_clear_congestion_window_wait(xprt);
405         __xprt_lock_write_next_cong(xprt);
406 }
407
408 /**
409  * xprt_request_get_cong - Request congestion control credits
410  * @xprt: pointer to transport
411  * @req: pointer to RPC request
412  *
413  * Useful for transports that require congestion control.
414  */
415 bool
416 xprt_request_get_cong(struct rpc_xprt *xprt, struct rpc_rqst *req)
417 {
418         bool ret = false;
419
420         if (req->rq_cong)
421                 return true;
422         spin_lock_bh(&xprt->transport_lock);
423         ret = __xprt_get_cong(xprt, req) != 0;
424         spin_unlock_bh(&xprt->transport_lock);
425         return ret;
426 }
427 EXPORT_SYMBOL_GPL(xprt_request_get_cong);
428
429 /**
430  * xprt_release_rqst_cong - housekeeping when request is complete
431  * @task: RPC request that recently completed
432  *
433  * Useful for transports that require congestion control.
434  */
435 void xprt_release_rqst_cong(struct rpc_task *task)
436 {
437         struct rpc_rqst *req = task->tk_rqstp;
438
439         __xprt_put_cong(req->rq_xprt, req);
440 }
441 EXPORT_SYMBOL_GPL(xprt_release_rqst_cong);
442
443 /*
444  * Clear the congestion window wait flag and wake up the next
445  * entry on xprt->sending
446  */
447 static void
448 xprt_clear_congestion_window_wait(struct rpc_xprt *xprt)
449 {
450         if (test_and_clear_bit(XPRT_CWND_WAIT, &xprt->state)) {
451                 spin_lock_bh(&xprt->transport_lock);
452                 __xprt_lock_write_next_cong(xprt);
453                 spin_unlock_bh(&xprt->transport_lock);
454         }
455 }
456
457 /**
458  * xprt_adjust_cwnd - adjust transport congestion window
459  * @xprt: pointer to xprt
460  * @task: recently completed RPC request used to adjust window
461  * @result: result code of completed RPC request
462  *
463  * The transport code maintains an estimate on the maximum number of out-
464  * standing RPC requests, using a smoothed version of the congestion
465  * avoidance implemented in 44BSD. This is basically the Van Jacobson
466  * congestion algorithm: If a retransmit occurs, the congestion window is
467  * halved; otherwise, it is incremented by 1/cwnd when
468  *
469  *      -       a reply is received and
470  *      -       a full number of requests are outstanding and
471  *      -       the congestion window hasn't been updated recently.
472  */
473 void xprt_adjust_cwnd(struct rpc_xprt *xprt, struct rpc_task *task, int result)
474 {
475         struct rpc_rqst *req = task->tk_rqstp;
476         unsigned long cwnd = xprt->cwnd;
477
478         if (result >= 0 && cwnd <= xprt->cong) {
479                 /* The (cwnd >> 1) term makes sure
480                  * the result gets rounded properly. */
481                 cwnd += (RPC_CWNDSCALE * RPC_CWNDSCALE + (cwnd >> 1)) / cwnd;
482                 if (cwnd > RPC_MAXCWND(xprt))
483                         cwnd = RPC_MAXCWND(xprt);
484                 __xprt_lock_write_next_cong(xprt);
485         } else if (result == -ETIMEDOUT) {
486                 cwnd >>= 1;
487                 if (cwnd < RPC_CWNDSCALE)
488                         cwnd = RPC_CWNDSCALE;
489         }
490         dprintk("RPC:       cong %ld, cwnd was %ld, now %ld\n",
491                         xprt->cong, xprt->cwnd, cwnd);
492         xprt->cwnd = cwnd;
493         __xprt_put_cong(xprt, req);
494 }
495 EXPORT_SYMBOL_GPL(xprt_adjust_cwnd);
496
497 /**
498  * xprt_wake_pending_tasks - wake all tasks on a transport's pending queue
499  * @xprt: transport with waiting tasks
500  * @status: result code to plant in each task before waking it
501  *
502  */
503 void xprt_wake_pending_tasks(struct rpc_xprt *xprt, int status)
504 {
505         if (status < 0)
506                 rpc_wake_up_status(&xprt->pending, status);
507         else
508                 rpc_wake_up(&xprt->pending);
509 }
510 EXPORT_SYMBOL_GPL(xprt_wake_pending_tasks);
511
512 /**
513  * xprt_wait_for_buffer_space - wait for transport output buffer to clear
514  * @xprt: transport
515  *
516  * Note that we only set the timer for the case of RPC_IS_SOFT(), since
517  * we don't in general want to force a socket disconnection due to
518  * an incomplete RPC call transmission.
519  */
520 void xprt_wait_for_buffer_space(struct rpc_xprt *xprt)
521 {
522         set_bit(XPRT_WRITE_SPACE, &xprt->state);
523 }
524 EXPORT_SYMBOL_GPL(xprt_wait_for_buffer_space);
525
526 static bool
527 xprt_clear_write_space_locked(struct rpc_xprt *xprt)
528 {
529         if (test_and_clear_bit(XPRT_WRITE_SPACE, &xprt->state)) {
530                 __xprt_lock_write_next(xprt);
531                 dprintk("RPC:       write space: waking waiting task on "
532                                 "xprt %p\n", xprt);
533                 return true;
534         }
535         return false;
536 }
537
538 /**
539  * xprt_write_space - wake the task waiting for transport output buffer space
540  * @xprt: transport with waiting tasks
541  *
542  * Can be called in a soft IRQ context, so xprt_write_space never sleeps.
543  */
544 bool xprt_write_space(struct rpc_xprt *xprt)
545 {
546         bool ret;
547
548         if (!test_bit(XPRT_WRITE_SPACE, &xprt->state))
549                 return false;
550         spin_lock_bh(&xprt->transport_lock);
551         ret = xprt_clear_write_space_locked(xprt);
552         spin_unlock_bh(&xprt->transport_lock);
553         return ret;
554 }
555 EXPORT_SYMBOL_GPL(xprt_write_space);
556
557 /**
558  * xprt_set_retrans_timeout_def - set a request's retransmit timeout
559  * @task: task whose timeout is to be set
560  *
561  * Set a request's retransmit timeout based on the transport's
562  * default timeout parameters.  Used by transports that don't adjust
563  * the retransmit timeout based on round-trip time estimation.
564  */
565 void xprt_set_retrans_timeout_def(struct rpc_task *task)
566 {
567         task->tk_timeout = task->tk_rqstp->rq_timeout;
568 }
569 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_def);
570
571 /**
572  * xprt_set_retrans_timeout_rtt - set a request's retransmit timeout
573  * @task: task whose timeout is to be set
574  *
575  * Set a request's retransmit timeout using the RTT estimator.
576  */
577 void xprt_set_retrans_timeout_rtt(struct rpc_task *task)
578 {
579         int timer = task->tk_msg.rpc_proc->p_timer;
580         struct rpc_clnt *clnt = task->tk_client;
581         struct rpc_rtt *rtt = clnt->cl_rtt;
582         struct rpc_rqst *req = task->tk_rqstp;
583         unsigned long max_timeout = clnt->cl_timeout->to_maxval;
584
585         task->tk_timeout = rpc_calc_rto(rtt, timer);
586         task->tk_timeout <<= rpc_ntimeo(rtt, timer) + req->rq_retries;
587         if (task->tk_timeout > max_timeout || task->tk_timeout == 0)
588                 task->tk_timeout = max_timeout;
589 }
590 EXPORT_SYMBOL_GPL(xprt_set_retrans_timeout_rtt);
591
592 static void xprt_reset_majortimeo(struct rpc_rqst *req)
593 {
594         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
595
596         req->rq_majortimeo = req->rq_timeout;
597         if (to->to_exponential)
598                 req->rq_majortimeo <<= to->to_retries;
599         else
600                 req->rq_majortimeo += to->to_increment * to->to_retries;
601         if (req->rq_majortimeo > to->to_maxval || req->rq_majortimeo == 0)
602                 req->rq_majortimeo = to->to_maxval;
603         req->rq_majortimeo += jiffies;
604 }
605
606 /**
607  * xprt_adjust_timeout - adjust timeout values for next retransmit
608  * @req: RPC request containing parameters to use for the adjustment
609  *
610  */
611 int xprt_adjust_timeout(struct rpc_rqst *req)
612 {
613         struct rpc_xprt *xprt = req->rq_xprt;
614         const struct rpc_timeout *to = req->rq_task->tk_client->cl_timeout;
615         int status = 0;
616
617         if (time_before(jiffies, req->rq_majortimeo)) {
618                 if (to->to_exponential)
619                         req->rq_timeout <<= 1;
620                 else
621                         req->rq_timeout += to->to_increment;
622                 if (to->to_maxval && req->rq_timeout >= to->to_maxval)
623                         req->rq_timeout = to->to_maxval;
624                 req->rq_retries++;
625         } else {
626                 req->rq_timeout = to->to_initval;
627                 req->rq_retries = 0;
628                 xprt_reset_majortimeo(req);
629                 /* Reset the RTT counters == "slow start" */
630                 spin_lock_bh(&xprt->transport_lock);
631                 rpc_init_rtt(req->rq_task->tk_client->cl_rtt, to->to_initval);
632                 spin_unlock_bh(&xprt->transport_lock);
633                 status = -ETIMEDOUT;
634         }
635
636         if (req->rq_timeout == 0) {
637                 printk(KERN_WARNING "xprt_adjust_timeout: rq_timeout = 0!\n");
638                 req->rq_timeout = 5 * HZ;
639         }
640         return status;
641 }
642
643 static void xprt_autoclose(struct work_struct *work)
644 {
645         struct rpc_xprt *xprt =
646                 container_of(work, struct rpc_xprt, task_cleanup);
647
648         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
649         xprt->ops->close(xprt);
650         xprt_release_write(xprt, NULL);
651         wake_up_bit(&xprt->state, XPRT_LOCKED);
652 }
653
654 /**
655  * xprt_disconnect_done - mark a transport as disconnected
656  * @xprt: transport to flag for disconnect
657  *
658  */
659 void xprt_disconnect_done(struct rpc_xprt *xprt)
660 {
661         dprintk("RPC:       disconnected transport %p\n", xprt);
662         spin_lock_bh(&xprt->transport_lock);
663         xprt_clear_connected(xprt);
664         xprt_clear_write_space_locked(xprt);
665         xprt_wake_pending_tasks(xprt, -EAGAIN);
666         spin_unlock_bh(&xprt->transport_lock);
667 }
668 EXPORT_SYMBOL_GPL(xprt_disconnect_done);
669
670 /**
671  * xprt_force_disconnect - force a transport to disconnect
672  * @xprt: transport to disconnect
673  *
674  */
675 void xprt_force_disconnect(struct rpc_xprt *xprt)
676 {
677         /* Don't race with the test_bit() in xprt_clear_locked() */
678         spin_lock_bh(&xprt->transport_lock);
679         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
680         /* Try to schedule an autoclose RPC call */
681         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
682                 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
683         xprt_wake_pending_tasks(xprt, -EAGAIN);
684         spin_unlock_bh(&xprt->transport_lock);
685 }
686 EXPORT_SYMBOL_GPL(xprt_force_disconnect);
687
688 static unsigned int
689 xprt_connect_cookie(struct rpc_xprt *xprt)
690 {
691         return READ_ONCE(xprt->connect_cookie);
692 }
693
694 static bool
695 xprt_request_retransmit_after_disconnect(struct rpc_task *task)
696 {
697         struct rpc_rqst *req = task->tk_rqstp;
698         struct rpc_xprt *xprt = req->rq_xprt;
699
700         return req->rq_connect_cookie != xprt_connect_cookie(xprt) ||
701                 !xprt_connected(xprt);
702 }
703
704 /**
705  * xprt_conditional_disconnect - force a transport to disconnect
706  * @xprt: transport to disconnect
707  * @cookie: 'connection cookie'
708  *
709  * This attempts to break the connection if and only if 'cookie' matches
710  * the current transport 'connection cookie'. It ensures that we don't
711  * try to break the connection more than once when we need to retransmit
712  * a batch of RPC requests.
713  *
714  */
715 void xprt_conditional_disconnect(struct rpc_xprt *xprt, unsigned int cookie)
716 {
717         /* Don't race with the test_bit() in xprt_clear_locked() */
718         spin_lock_bh(&xprt->transport_lock);
719         if (cookie != xprt->connect_cookie)
720                 goto out;
721         if (test_bit(XPRT_CLOSING, &xprt->state))
722                 goto out;
723         set_bit(XPRT_CLOSE_WAIT, &xprt->state);
724         /* Try to schedule an autoclose RPC call */
725         if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
726                 queue_work(xprtiod_workqueue, &xprt->task_cleanup);
727         xprt_wake_pending_tasks(xprt, -EAGAIN);
728 out:
729         spin_unlock_bh(&xprt->transport_lock);
730 }
731
732 static bool
733 xprt_has_timer(const struct rpc_xprt *xprt)
734 {
735         return xprt->idle_timeout != 0;
736 }
737
738 static void
739 xprt_schedule_autodisconnect(struct rpc_xprt *xprt)
740         __must_hold(&xprt->transport_lock)
741 {
742         if (RB_EMPTY_ROOT(&xprt->recv_queue) && xprt_has_timer(xprt))
743                 mod_timer(&xprt->timer, xprt->last_used + xprt->idle_timeout);
744 }
745
746 static void
747 xprt_init_autodisconnect(struct timer_list *t)
748 {
749         struct rpc_xprt *xprt = from_timer(xprt, t, timer);
750
751         spin_lock(&xprt->transport_lock);
752         if (!RB_EMPTY_ROOT(&xprt->recv_queue))
753                 goto out_abort;
754         /* Reset xprt->last_used to avoid connect/autodisconnect cycling */
755         xprt->last_used = jiffies;
756         if (test_and_set_bit(XPRT_LOCKED, &xprt->state))
757                 goto out_abort;
758         spin_unlock(&xprt->transport_lock);
759         queue_work(xprtiod_workqueue, &xprt->task_cleanup);
760         return;
761 out_abort:
762         spin_unlock(&xprt->transport_lock);
763 }
764
765 bool xprt_lock_connect(struct rpc_xprt *xprt,
766                 struct rpc_task *task,
767                 void *cookie)
768 {
769         bool ret = false;
770
771         spin_lock_bh(&xprt->transport_lock);
772         if (!test_bit(XPRT_LOCKED, &xprt->state))
773                 goto out;
774         if (xprt->snd_task != task)
775                 goto out;
776         xprt->snd_task = cookie;
777         ret = true;
778 out:
779         spin_unlock_bh(&xprt->transport_lock);
780         return ret;
781 }
782
783 void xprt_unlock_connect(struct rpc_xprt *xprt, void *cookie)
784 {
785         spin_lock_bh(&xprt->transport_lock);
786         if (xprt->snd_task != cookie)
787                 goto out;
788         if (!test_bit(XPRT_LOCKED, &xprt->state))
789                 goto out;
790         xprt->snd_task =NULL;
791         xprt->ops->release_xprt(xprt, NULL);
792         xprt_schedule_autodisconnect(xprt);
793 out:
794         spin_unlock_bh(&xprt->transport_lock);
795         wake_up_bit(&xprt->state, XPRT_LOCKED);
796 }
797
798 /**
799  * xprt_connect - schedule a transport connect operation
800  * @task: RPC task that is requesting the connect
801  *
802  */
803 void xprt_connect(struct rpc_task *task)
804 {
805         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
806
807         dprintk("RPC: %5u xprt_connect xprt %p %s connected\n", task->tk_pid,
808                         xprt, (xprt_connected(xprt) ? "is" : "is not"));
809
810         if (!xprt_bound(xprt)) {
811                 task->tk_status = -EAGAIN;
812                 return;
813         }
814         if (!xprt_lock_write(xprt, task))
815                 return;
816
817         if (test_and_clear_bit(XPRT_CLOSE_WAIT, &xprt->state))
818                 xprt->ops->close(xprt);
819
820         if (!xprt_connected(xprt)) {
821                 task->tk_timeout = task->tk_rqstp->rq_timeout;
822                 task->tk_rqstp->rq_connect_cookie = xprt->connect_cookie;
823                 rpc_sleep_on(&xprt->pending, task, xprt_connect_status);
824
825                 if (test_bit(XPRT_CLOSING, &xprt->state))
826                         return;
827                 if (xprt_test_and_set_connecting(xprt))
828                         return;
829                 xprt->stat.connect_start = jiffies;
830                 xprt->ops->connect(xprt, task);
831         }
832         xprt_release_write(xprt, task);
833 }
834
835 static void xprt_connect_status(struct rpc_task *task)
836 {
837         struct rpc_xprt *xprt = task->tk_rqstp->rq_xprt;
838
839         if (task->tk_status == 0) {
840                 xprt->stat.connect_count++;
841                 xprt->stat.connect_time += (long)jiffies - xprt->stat.connect_start;
842                 dprintk("RPC: %5u xprt_connect_status: connection established\n",
843                                 task->tk_pid);
844                 return;
845         }
846
847         switch (task->tk_status) {
848         case -ECONNREFUSED:
849         case -ECONNRESET:
850         case -ECONNABORTED:
851         case -ENETUNREACH:
852         case -EHOSTUNREACH:
853         case -EPIPE:
854         case -EAGAIN:
855                 dprintk("RPC: %5u xprt_connect_status: retrying\n", task->tk_pid);
856                 break;
857         case -ETIMEDOUT:
858                 dprintk("RPC: %5u xprt_connect_status: connect attempt timed "
859                                 "out\n", task->tk_pid);
860                 break;
861         default:
862                 dprintk("RPC: %5u xprt_connect_status: error %d connecting to "
863                                 "server %s\n", task->tk_pid, -task->tk_status,
864                                 xprt->servername);
865                 task->tk_status = -EIO;
866         }
867 }
868
869 enum xprt_xid_rb_cmp {
870         XID_RB_EQUAL,
871         XID_RB_LEFT,
872         XID_RB_RIGHT,
873 };
874 static enum xprt_xid_rb_cmp
875 xprt_xid_cmp(__be32 xid1, __be32 xid2)
876 {
877         if (xid1 == xid2)
878                 return XID_RB_EQUAL;
879         if ((__force u32)xid1 < (__force u32)xid2)
880                 return XID_RB_LEFT;
881         return XID_RB_RIGHT;
882 }
883
884 static struct rpc_rqst *
885 xprt_request_rb_find(struct rpc_xprt *xprt, __be32 xid)
886 {
887         struct rb_node *n = xprt->recv_queue.rb_node;
888         struct rpc_rqst *req;
889
890         while (n != NULL) {
891                 req = rb_entry(n, struct rpc_rqst, rq_recv);
892                 switch (xprt_xid_cmp(xid, req->rq_xid)) {
893                 case XID_RB_LEFT:
894                         n = n->rb_left;
895                         break;
896                 case XID_RB_RIGHT:
897                         n = n->rb_right;
898                         break;
899                 case XID_RB_EQUAL:
900                         return req;
901                 }
902         }
903         return NULL;
904 }
905
906 static void
907 xprt_request_rb_insert(struct rpc_xprt *xprt, struct rpc_rqst *new)
908 {
909         struct rb_node **p = &xprt->recv_queue.rb_node;
910         struct rb_node *n = NULL;
911         struct rpc_rqst *req;
912
913         while (*p != NULL) {
914                 n = *p;
915                 req = rb_entry(n, struct rpc_rqst, rq_recv);
916                 switch(xprt_xid_cmp(new->rq_xid, req->rq_xid)) {
917                 case XID_RB_LEFT:
918                         p = &n->rb_left;
919                         break;
920                 case XID_RB_RIGHT:
921                         p = &n->rb_right;
922                         break;
923                 case XID_RB_EQUAL:
924                         WARN_ON_ONCE(new != req);
925                         return;
926                 }
927         }
928         rb_link_node(&new->rq_recv, n, p);
929         rb_insert_color(&new->rq_recv, &xprt->recv_queue);
930 }
931
932 static void
933 xprt_request_rb_remove(struct rpc_xprt *xprt, struct rpc_rqst *req)
934 {
935         rb_erase(&req->rq_recv, &xprt->recv_queue);
936 }
937
938 /**
939  * xprt_lookup_rqst - find an RPC request corresponding to an XID
940  * @xprt: transport on which the original request was transmitted
941  * @xid: RPC XID of incoming reply
942  *
943  * Caller holds xprt->queue_lock.
944  */
945 struct rpc_rqst *xprt_lookup_rqst(struct rpc_xprt *xprt, __be32 xid)
946 {
947         struct rpc_rqst *entry;
948
949         entry = xprt_request_rb_find(xprt, xid);
950         if (entry != NULL) {
951                 trace_xprt_lookup_rqst(xprt, xid, 0);
952                 entry->rq_rtt = ktime_sub(ktime_get(), entry->rq_xtime);
953                 return entry;
954         }
955
956         dprintk("RPC:       xprt_lookup_rqst did not find xid %08x\n",
957                         ntohl(xid));
958         trace_xprt_lookup_rqst(xprt, xid, -ENOENT);
959         xprt->stat.bad_xids++;
960         return NULL;
961 }
962 EXPORT_SYMBOL_GPL(xprt_lookup_rqst);
963
964 static bool
965 xprt_is_pinned_rqst(struct rpc_rqst *req)
966 {
967         return atomic_read(&req->rq_pin) != 0;
968 }
969
970 /**
971  * xprt_pin_rqst - Pin a request on the transport receive list
972  * @req: Request to pin
973  *
974  * Caller must ensure this is atomic with the call to xprt_lookup_rqst()
975  * so should be holding the xprt receive lock.
976  */
977 void xprt_pin_rqst(struct rpc_rqst *req)
978 {
979         atomic_inc(&req->rq_pin);
980 }
981 EXPORT_SYMBOL_GPL(xprt_pin_rqst);
982
983 /**
984  * xprt_unpin_rqst - Unpin a request on the transport receive list
985  * @req: Request to pin
986  *
987  * Caller should be holding the xprt receive lock.
988  */
989 void xprt_unpin_rqst(struct rpc_rqst *req)
990 {
991         if (!test_bit(RPC_TASK_MSG_PIN_WAIT, &req->rq_task->tk_runstate)) {
992                 atomic_dec(&req->rq_pin);
993                 return;
994         }
995         if (atomic_dec_and_test(&req->rq_pin))
996                 wake_up_var(&req->rq_pin);
997 }
998 EXPORT_SYMBOL_GPL(xprt_unpin_rqst);
999
1000 static void xprt_wait_on_pinned_rqst(struct rpc_rqst *req)
1001 {
1002         wait_var_event(&req->rq_pin, !xprt_is_pinned_rqst(req));
1003 }
1004
1005 static bool
1006 xprt_request_data_received(struct rpc_task *task)
1007 {
1008         return !test_bit(RPC_TASK_NEED_RECV, &task->tk_runstate) &&
1009                 READ_ONCE(task->tk_rqstp->rq_reply_bytes_recvd) != 0;
1010 }
1011
1012 static bool
1013 xprt_request_need_enqueue_receive(struct rpc_task *task, struct rpc_rqst *req)
1014 {
1015         return !test_bit(RPC_TASK_NEED_RECV, &task->tk_runstate) &&
1016                 READ_ONCE(task->tk_rqstp->rq_reply_bytes_recvd) == 0;
1017 }
1018
1019 /**
1020  * xprt_request_enqueue_receive - Add an request to the receive queue
1021  * @task: RPC task
1022  *
1023  */
1024 void
1025 xprt_request_enqueue_receive(struct rpc_task *task)
1026 {
1027         struct rpc_rqst *req = task->tk_rqstp;
1028         struct rpc_xprt *xprt = req->rq_xprt;
1029
1030         if (!xprt_request_need_enqueue_receive(task, req))
1031                 return;
1032         spin_lock(&xprt->queue_lock);
1033
1034         /* Update the softirq receive buffer */
1035         memcpy(&req->rq_private_buf, &req->rq_rcv_buf,
1036                         sizeof(req->rq_private_buf));
1037
1038         /* Add request to the receive list */
1039         xprt_request_rb_insert(xprt, req);
1040         set_bit(RPC_TASK_NEED_RECV, &task->tk_runstate);
1041         spin_unlock(&xprt->queue_lock);
1042
1043         xprt_reset_majortimeo(req);
1044         /* Turn off autodisconnect */
1045         del_singleshot_timer_sync(&xprt->timer);
1046 }
1047
1048 /**
1049  * xprt_request_dequeue_receive_locked - Remove a request from the receive queue
1050  * @task: RPC task
1051  *
1052  * Caller must hold xprt->queue_lock.
1053  */
1054 static void
1055 xprt_request_dequeue_receive_locked(struct rpc_task *task)
1056 {
1057         struct rpc_rqst *req = task->tk_rqstp;
1058
1059         if (test_and_clear_bit(RPC_TASK_NEED_RECV, &task->tk_runstate))
1060                 xprt_request_rb_remove(req->rq_xprt, req);
1061 }
1062
1063 /**
1064  * xprt_update_rtt - Update RPC RTT statistics
1065  * @task: RPC request that recently completed
1066  *
1067  * Caller holds xprt->queue_lock.
1068  */
1069 void xprt_update_rtt(struct rpc_task *task)
1070 {
1071         struct rpc_rqst *req = task->tk_rqstp;
1072         struct rpc_rtt *rtt = task->tk_client->cl_rtt;
1073         unsigned int timer = task->tk_msg.rpc_proc->p_timer;
1074         long m = usecs_to_jiffies(ktime_to_us(req->rq_rtt));
1075
1076         if (timer) {
1077                 if (req->rq_ntrans == 1)
1078                         rpc_update_rtt(rtt, timer, m);
1079                 rpc_set_timeo(rtt, timer, req->rq_ntrans - 1);
1080         }
1081 }
1082 EXPORT_SYMBOL_GPL(xprt_update_rtt);
1083
1084 /**
1085  * xprt_complete_rqst - called when reply processing is complete
1086  * @task: RPC request that recently completed
1087  * @copied: actual number of bytes received from the transport
1088  *
1089  * Caller holds xprt->queue_lock.
1090  */
1091 void xprt_complete_rqst(struct rpc_task *task, int copied)
1092 {
1093         struct rpc_rqst *req = task->tk_rqstp;
1094         struct rpc_xprt *xprt = req->rq_xprt;
1095
1096         dprintk("RPC: %5u xid %08x complete (%d bytes received)\n",
1097                         task->tk_pid, ntohl(req->rq_xid), copied);
1098         trace_xprt_complete_rqst(xprt, req->rq_xid, copied);
1099
1100         xprt->stat.recvs++;
1101
1102         req->rq_private_buf.len = copied;
1103         /* Ensure all writes are done before we update */
1104         /* req->rq_reply_bytes_recvd */
1105         smp_wmb();
1106         req->rq_reply_bytes_recvd = copied;
1107         xprt_request_dequeue_receive_locked(task);
1108         rpc_wake_up_queued_task(&xprt->pending, task);
1109 }
1110 EXPORT_SYMBOL_GPL(xprt_complete_rqst);
1111
1112 static void xprt_timer(struct rpc_task *task)
1113 {
1114         struct rpc_rqst *req = task->tk_rqstp;
1115         struct rpc_xprt *xprt = req->rq_xprt;
1116
1117         if (task->tk_status != -ETIMEDOUT)
1118                 return;
1119
1120         trace_xprt_timer(xprt, req->rq_xid, task->tk_status);
1121         if (!req->rq_reply_bytes_recvd) {
1122                 if (xprt->ops->timer)
1123                         xprt->ops->timer(xprt, task);
1124         } else
1125                 task->tk_status = 0;
1126 }
1127
1128 /**
1129  * xprt_request_wait_receive - wait for the reply to an RPC request
1130  * @task: RPC task about to send a request
1131  *
1132  */
1133 void xprt_request_wait_receive(struct rpc_task *task)
1134 {
1135         struct rpc_rqst *req = task->tk_rqstp;
1136         struct rpc_xprt *xprt = req->rq_xprt;
1137
1138         if (!test_bit(RPC_TASK_NEED_RECV, &task->tk_runstate))
1139                 return;
1140         /*
1141          * Sleep on the pending queue if we're expecting a reply.
1142          * The spinlock ensures atomicity between the test of
1143          * req->rq_reply_bytes_recvd, and the call to rpc_sleep_on().
1144          */
1145         spin_lock(&xprt->queue_lock);
1146         if (test_bit(RPC_TASK_NEED_RECV, &task->tk_runstate)) {
1147                 xprt->ops->set_retrans_timeout(task);
1148                 rpc_sleep_on(&xprt->pending, task, xprt_timer);
1149                 /*
1150                  * Send an extra queue wakeup call if the
1151                  * connection was dropped in case the call to
1152                  * rpc_sleep_on() raced.
1153                  */
1154                 if (xprt_request_retransmit_after_disconnect(task))
1155                         rpc_wake_up_queued_task_set_status(&xprt->pending,
1156                                         task, -ENOTCONN);
1157         }
1158         spin_unlock(&xprt->queue_lock);
1159 }
1160
1161 static bool
1162 xprt_request_need_enqueue_transmit(struct rpc_task *task, struct rpc_rqst *req)
1163 {
1164         return !test_bit(RPC_TASK_NEED_XMIT, &task->tk_runstate);
1165 }
1166
1167 /**
1168  * xprt_request_enqueue_transmit - queue a task for transmission
1169  * @task: pointer to rpc_task
1170  *
1171  * Add a task to the transmission queue.
1172  */
1173 void
1174 xprt_request_enqueue_transmit(struct rpc_task *task)
1175 {
1176         struct rpc_rqst *pos, *req = task->tk_rqstp;
1177         struct rpc_xprt *xprt = req->rq_xprt;
1178
1179         if (xprt_request_need_enqueue_transmit(task, req)) {
1180                 spin_lock(&xprt->queue_lock);
1181                 /*
1182                  * Requests that carry congestion control credits are added
1183                  * to the head of the list to avoid starvation issues.
1184                  */
1185                 if (req->rq_cong) {
1186                         xprt_clear_congestion_window_wait(xprt);
1187                         list_for_each_entry(pos, &xprt->xmit_queue, rq_xmit) {
1188                                 if (pos->rq_cong)
1189                                         continue;
1190                                 /* Note: req is added _before_ pos */
1191                                 list_add_tail(&req->rq_xmit, &pos->rq_xmit);
1192                                 INIT_LIST_HEAD(&req->rq_xmit2);
1193                                 goto out;
1194                         }
1195                 } else if (RPC_IS_SWAPPER(task)) {
1196                         list_for_each_entry(pos, &xprt->xmit_queue, rq_xmit) {
1197                                 if (pos->rq_cong || pos->rq_bytes_sent)
1198                                         continue;
1199                                 if (RPC_IS_SWAPPER(pos->rq_task))
1200                                         continue;
1201                                 /* Note: req is added _before_ pos */
1202                                 list_add_tail(&req->rq_xmit, &pos->rq_xmit);
1203                                 INIT_LIST_HEAD(&req->rq_xmit2);
1204                                 goto out;
1205                         }
1206                 } else {
1207                         list_for_each_entry(pos, &xprt->xmit_queue, rq_xmit) {
1208                                 if (pos->rq_task->tk_owner != task->tk_owner)
1209                                         continue;
1210                                 list_add_tail(&req->rq_xmit2, &pos->rq_xmit2);
1211                                 INIT_LIST_HEAD(&req->rq_xmit);
1212                                 goto out;
1213                         }
1214                 }
1215                 list_add_tail(&req->rq_xmit, &xprt->xmit_queue);
1216                 INIT_LIST_HEAD(&req->rq_xmit2);
1217 out:
1218                 set_bit(RPC_TASK_NEED_XMIT, &task->tk_runstate);
1219                 spin_unlock(&xprt->queue_lock);
1220         }
1221 }
1222
1223 /**
1224  * xprt_request_dequeue_transmit_locked - remove a task from the transmission queue
1225  * @task: pointer to rpc_task
1226  *
1227  * Remove a task from the transmission queue
1228  * Caller must hold xprt->queue_lock
1229  */
1230 static void
1231 xprt_request_dequeue_transmit_locked(struct rpc_task *task)
1232 {
1233         struct rpc_rqst *req = task->tk_rqstp;
1234
1235         if (!test_and_clear_bit(RPC_TASK_NEED_XMIT, &task->tk_runstate))
1236                 return;
1237         if (!list_empty(&req->rq_xmit)) {
1238                 list_del(&req->rq_xmit);
1239                 if (!list_empty(&req->rq_xmit2)) {
1240                         struct rpc_rqst *next = list_first_entry(&req->rq_xmit2,
1241                                         struct rpc_rqst, rq_xmit2);
1242                         list_del(&req->rq_xmit2);
1243                         list_add_tail(&next->rq_xmit, &next->rq_xprt->xmit_queue);
1244                 }
1245         } else
1246                 list_del(&req->rq_xmit2);
1247 }
1248
1249 /**
1250  * xprt_request_dequeue_transmit - remove a task from the transmission queue
1251  * @task: pointer to rpc_task
1252  *
1253  * Remove a task from the transmission queue
1254  */
1255 static void
1256 xprt_request_dequeue_transmit(struct rpc_task *task)
1257 {
1258         struct rpc_rqst *req = task->tk_rqstp;
1259         struct rpc_xprt *xprt = req->rq_xprt;
1260
1261         spin_lock(&xprt->queue_lock);
1262         xprt_request_dequeue_transmit_locked(task);
1263         spin_unlock(&xprt->queue_lock);
1264 }
1265
1266 /**
1267  * xprt_request_need_retransmit - Test if a task needs retransmission
1268  * @task: pointer to rpc_task
1269  *
1270  * Test for whether a connection breakage requires the task to retransmit
1271  */
1272 bool
1273 xprt_request_need_retransmit(struct rpc_task *task)
1274 {
1275         return xprt_request_retransmit_after_disconnect(task);
1276 }
1277
1278 /**
1279  * xprt_prepare_transmit - reserve the transport before sending a request
1280  * @task: RPC task about to send a request
1281  *
1282  */
1283 bool xprt_prepare_transmit(struct rpc_task *task)
1284 {
1285         struct rpc_rqst *req = task->tk_rqstp;
1286         struct rpc_xprt *xprt = req->rq_xprt;
1287
1288         dprintk("RPC: %5u xprt_prepare_transmit\n", task->tk_pid);
1289
1290         if (!xprt_lock_write(xprt, task)) {
1291                 /* Race breaker: someone may have transmitted us */
1292                 if (!test_bit(RPC_TASK_NEED_XMIT, &task->tk_runstate))
1293                         rpc_wake_up_queued_task_set_status(&xprt->sending,
1294                                         task, 0);
1295                 return false;
1296
1297         }
1298         return true;
1299 }
1300
1301 void xprt_end_transmit(struct rpc_task *task)
1302 {
1303         xprt_release_write(task->tk_rqstp->rq_xprt, task);
1304 }
1305
1306 /**
1307  * xprt_request_transmit - send an RPC request on a transport
1308  * @req: pointer to request to transmit
1309  * @snd_task: RPC task that owns the transport lock
1310  *
1311  * This performs the transmission of a single request.
1312  * Note that if the request is not the same as snd_task, then it
1313  * does need to be pinned.
1314  * Returns '0' on success.
1315  */
1316 static int
1317 xprt_request_transmit(struct rpc_rqst *req, struct rpc_task *snd_task)
1318 {
1319         struct rpc_xprt *xprt = req->rq_xprt;
1320         struct rpc_task *task = req->rq_task;
1321         unsigned int connect_cookie;
1322         int is_retrans = RPC_WAS_SENT(task);
1323         int status;
1324
1325         dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen);
1326
1327         if (!req->rq_bytes_sent) {
1328                 if (xprt_request_data_received(task)) {
1329                         status = 0;
1330                         goto out_dequeue;
1331                 }
1332                 /* Verify that our message lies in the RPCSEC_GSS window */
1333                 if (rpcauth_xmit_need_reencode(task)) {
1334                         status = -EBADMSG;
1335                         goto out_dequeue;
1336                 }
1337         }
1338
1339         /*
1340          * Update req->rq_ntrans before transmitting to avoid races with
1341          * xprt_update_rtt(), which needs to know that it is recording a
1342          * reply to the first transmission.
1343          */
1344         req->rq_ntrans++;
1345
1346         connect_cookie = xprt->connect_cookie;
1347         status = xprt->ops->send_request(req);
1348         trace_xprt_transmit(xprt, req->rq_xid, status);
1349         if (status != 0) {
1350                 req->rq_ntrans--;
1351                 return status;
1352         }
1353
1354         if (is_retrans)
1355                 task->tk_client->cl_stats->rpcretrans++;
1356
1357         xprt_inject_disconnect(xprt);
1358
1359         dprintk("RPC: %5u xmit complete\n", task->tk_pid);
1360         task->tk_flags |= RPC_TASK_SENT;
1361         spin_lock_bh(&xprt->transport_lock);
1362
1363         xprt->stat.sends++;
1364         xprt->stat.req_u += xprt->stat.sends - xprt->stat.recvs;
1365         xprt->stat.bklog_u += xprt->backlog.qlen;
1366         xprt->stat.sending_u += xprt->sending.qlen;
1367         xprt->stat.pending_u += xprt->pending.qlen;
1368         spin_unlock_bh(&xprt->transport_lock);
1369
1370         req->rq_connect_cookie = connect_cookie;
1371 out_dequeue:
1372         xprt_request_dequeue_transmit(task);
1373         rpc_wake_up_queued_task_set_status(&xprt->sending, task, status);
1374         return status;
1375 }
1376
1377 /**
1378  * xprt_transmit - send an RPC request on a transport
1379  * @task: controlling RPC task
1380  *
1381  * Attempts to drain the transmit queue. On exit, either the transport
1382  * signalled an error that needs to be handled before transmission can
1383  * resume, or @task finished transmitting, and detected that it already
1384  * received a reply.
1385  */
1386 void
1387 xprt_transmit(struct rpc_task *task)
1388 {
1389         struct rpc_rqst *next, *req = task->tk_rqstp;
1390         struct rpc_xprt *xprt = req->rq_xprt;
1391         int status;
1392
1393         spin_lock(&xprt->queue_lock);
1394         while (!list_empty(&xprt->xmit_queue)) {
1395                 next = list_first_entry(&xprt->xmit_queue,
1396                                 struct rpc_rqst, rq_xmit);
1397                 xprt_pin_rqst(next);
1398                 spin_unlock(&xprt->queue_lock);
1399                 status = xprt_request_transmit(next, task);
1400                 if (status == -EBADMSG && next != req)
1401                         status = 0;
1402                 cond_resched();
1403                 spin_lock(&xprt->queue_lock);
1404                 xprt_unpin_rqst(next);
1405                 if (status == 0) {
1406                         if (!xprt_request_data_received(task) ||
1407                             test_bit(RPC_TASK_NEED_XMIT, &task->tk_runstate))
1408                                 continue;
1409                 } else if (test_bit(RPC_TASK_NEED_XMIT, &task->tk_runstate))
1410                         task->tk_status = status;
1411                 break;
1412         }
1413         spin_unlock(&xprt->queue_lock);
1414 }
1415
1416 static void xprt_add_backlog(struct rpc_xprt *xprt, struct rpc_task *task)
1417 {
1418         set_bit(XPRT_CONGESTED, &xprt->state);
1419         rpc_sleep_on(&xprt->backlog, task, NULL);
1420 }
1421
1422 static void xprt_wake_up_backlog(struct rpc_xprt *xprt)
1423 {
1424         if (rpc_wake_up_next(&xprt->backlog) == NULL)
1425                 clear_bit(XPRT_CONGESTED, &xprt->state);
1426 }
1427
1428 static bool xprt_throttle_congested(struct rpc_xprt *xprt, struct rpc_task *task)
1429 {
1430         bool ret = false;
1431
1432         if (!test_bit(XPRT_CONGESTED, &xprt->state))
1433                 goto out;
1434         spin_lock(&xprt->reserve_lock);
1435         if (test_bit(XPRT_CONGESTED, &xprt->state)) {
1436                 rpc_sleep_on(&xprt->backlog, task, NULL);
1437                 ret = true;
1438         }
1439         spin_unlock(&xprt->reserve_lock);
1440 out:
1441         return ret;
1442 }
1443
1444 static struct rpc_rqst *xprt_dynamic_alloc_slot(struct rpc_xprt *xprt)
1445 {
1446         struct rpc_rqst *req = ERR_PTR(-EAGAIN);
1447
1448         if (xprt->num_reqs >= xprt->max_reqs)
1449                 goto out;
1450         ++xprt->num_reqs;
1451         spin_unlock(&xprt->reserve_lock);
1452         req = kzalloc(sizeof(struct rpc_rqst), GFP_NOFS);
1453         spin_lock(&xprt->reserve_lock);
1454         if (req != NULL)
1455                 goto out;
1456         --xprt->num_reqs;
1457         req = ERR_PTR(-ENOMEM);
1458 out:
1459         return req;
1460 }
1461
1462 static bool xprt_dynamic_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1463 {
1464         if (xprt->num_reqs > xprt->min_reqs) {
1465                 --xprt->num_reqs;
1466                 kfree(req);
1467                 return true;
1468         }
1469         return false;
1470 }
1471
1472 void xprt_alloc_slot(struct rpc_xprt *xprt, struct rpc_task *task)
1473 {
1474         struct rpc_rqst *req;
1475
1476         spin_lock(&xprt->reserve_lock);
1477         if (!list_empty(&xprt->free)) {
1478                 req = list_entry(xprt->free.next, struct rpc_rqst, rq_list);
1479                 list_del(&req->rq_list);
1480                 goto out_init_req;
1481         }
1482         req = xprt_dynamic_alloc_slot(xprt);
1483         if (!IS_ERR(req))
1484                 goto out_init_req;
1485         switch (PTR_ERR(req)) {
1486         case -ENOMEM:
1487                 dprintk("RPC:       dynamic allocation of request slot "
1488                                 "failed! Retrying\n");
1489                 task->tk_status = -ENOMEM;
1490                 break;
1491         case -EAGAIN:
1492                 xprt_add_backlog(xprt, task);
1493                 dprintk("RPC:       waiting for request slot\n");
1494                 /* fall through */
1495         default:
1496                 task->tk_status = -EAGAIN;
1497         }
1498         spin_unlock(&xprt->reserve_lock);
1499         return;
1500 out_init_req:
1501         xprt->stat.max_slots = max_t(unsigned int, xprt->stat.max_slots,
1502                                      xprt->num_reqs);
1503         spin_unlock(&xprt->reserve_lock);
1504
1505         task->tk_status = 0;
1506         task->tk_rqstp = req;
1507 }
1508 EXPORT_SYMBOL_GPL(xprt_alloc_slot);
1509
1510 void xprt_free_slot(struct rpc_xprt *xprt, struct rpc_rqst *req)
1511 {
1512         spin_lock(&xprt->reserve_lock);
1513         if (!xprt_dynamic_free_slot(xprt, req)) {
1514                 memset(req, 0, sizeof(*req));   /* mark unused */
1515                 list_add(&req->rq_list, &xprt->free);
1516         }
1517         xprt_wake_up_backlog(xprt);
1518         spin_unlock(&xprt->reserve_lock);
1519 }
1520 EXPORT_SYMBOL_GPL(xprt_free_slot);
1521
1522 static void xprt_free_all_slots(struct rpc_xprt *xprt)
1523 {
1524         struct rpc_rqst *req;
1525         while (!list_empty(&xprt->free)) {
1526                 req = list_first_entry(&xprt->free, struct rpc_rqst, rq_list);
1527                 list_del(&req->rq_list);
1528                 kfree(req);
1529         }
1530 }
1531
1532 struct rpc_xprt *xprt_alloc(struct net *net, size_t size,
1533                 unsigned int num_prealloc,
1534                 unsigned int max_alloc)
1535 {
1536         struct rpc_xprt *xprt;
1537         struct rpc_rqst *req;
1538         int i;
1539
1540         xprt = kzalloc(size, GFP_KERNEL);
1541         if (xprt == NULL)
1542                 goto out;
1543
1544         xprt_init(xprt, net);
1545
1546         for (i = 0; i < num_prealloc; i++) {
1547                 req = kzalloc(sizeof(struct rpc_rqst), GFP_KERNEL);
1548                 if (!req)
1549                         goto out_free;
1550                 list_add(&req->rq_list, &xprt->free);
1551         }
1552         if (max_alloc > num_prealloc)
1553                 xprt->max_reqs = max_alloc;
1554         else
1555                 xprt->max_reqs = num_prealloc;
1556         xprt->min_reqs = num_prealloc;
1557         xprt->num_reqs = num_prealloc;
1558
1559         return xprt;
1560
1561 out_free:
1562         xprt_free(xprt);
1563 out:
1564         return NULL;
1565 }
1566 EXPORT_SYMBOL_GPL(xprt_alloc);
1567
1568 void xprt_free(struct rpc_xprt *xprt)
1569 {
1570         put_net(xprt->xprt_net);
1571         xprt_free_all_slots(xprt);
1572         kfree_rcu(xprt, rcu);
1573 }
1574 EXPORT_SYMBOL_GPL(xprt_free);
1575
1576 static void
1577 xprt_init_connect_cookie(struct rpc_rqst *req, struct rpc_xprt *xprt)
1578 {
1579         req->rq_connect_cookie = xprt_connect_cookie(xprt) - 1;
1580 }
1581
1582 static __be32
1583 xprt_alloc_xid(struct rpc_xprt *xprt)
1584 {
1585         __be32 xid;
1586
1587         spin_lock(&xprt->reserve_lock);
1588         xid = (__force __be32)xprt->xid++;
1589         spin_unlock(&xprt->reserve_lock);
1590         return xid;
1591 }
1592
1593 static void
1594 xprt_init_xid(struct rpc_xprt *xprt)
1595 {
1596         xprt->xid = prandom_u32();
1597 }
1598
1599 static void
1600 xprt_request_init(struct rpc_task *task)
1601 {
1602         struct rpc_xprt *xprt = task->tk_xprt;
1603         struct rpc_rqst *req = task->tk_rqstp;
1604
1605         req->rq_timeout = task->tk_client->cl_timeout->to_initval;
1606         req->rq_task    = task;
1607         req->rq_xprt    = xprt;
1608         req->rq_buffer  = NULL;
1609         req->rq_xid     = xprt_alloc_xid(xprt);
1610         xprt_init_connect_cookie(req, xprt);
1611         req->rq_bytes_sent = 0;
1612         req->rq_snd_buf.len = 0;
1613         req->rq_snd_buf.buflen = 0;
1614         req->rq_rcv_buf.len = 0;
1615         req->rq_rcv_buf.buflen = 0;
1616         req->rq_release_snd_buf = NULL;
1617         xprt_reset_majortimeo(req);
1618         dprintk("RPC: %5u reserved req %p xid %08x\n", task->tk_pid,
1619                         req, ntohl(req->rq_xid));
1620 }
1621
1622 static void
1623 xprt_do_reserve(struct rpc_xprt *xprt, struct rpc_task *task)
1624 {
1625         xprt->ops->alloc_slot(xprt, task);
1626         if (task->tk_rqstp != NULL)
1627                 xprt_request_init(task);
1628 }
1629
1630 /**
1631  * xprt_reserve - allocate an RPC request slot
1632  * @task: RPC task requesting a slot allocation
1633  *
1634  * If the transport is marked as being congested, or if no more
1635  * slots are available, place the task on the transport's
1636  * backlog queue.
1637  */
1638 void xprt_reserve(struct rpc_task *task)
1639 {
1640         struct rpc_xprt *xprt = task->tk_xprt;
1641
1642         task->tk_status = 0;
1643         if (task->tk_rqstp != NULL)
1644                 return;
1645
1646         task->tk_timeout = 0;
1647         task->tk_status = -EAGAIN;
1648         if (!xprt_throttle_congested(xprt, task))
1649                 xprt_do_reserve(xprt, task);
1650 }
1651
1652 /**
1653  * xprt_retry_reserve - allocate an RPC request slot
1654  * @task: RPC task requesting a slot allocation
1655  *
1656  * If no more slots are available, place the task on the transport's
1657  * backlog queue.
1658  * Note that the only difference with xprt_reserve is that we now
1659  * ignore the value of the XPRT_CONGESTED flag.
1660  */
1661 void xprt_retry_reserve(struct rpc_task *task)
1662 {
1663         struct rpc_xprt *xprt = task->tk_xprt;
1664
1665         task->tk_status = 0;
1666         if (task->tk_rqstp != NULL)
1667                 return;
1668
1669         task->tk_timeout = 0;
1670         task->tk_status = -EAGAIN;
1671         xprt_do_reserve(xprt, task);
1672 }
1673
1674 static void
1675 xprt_request_dequeue_all(struct rpc_task *task, struct rpc_rqst *req)
1676 {
1677         struct rpc_xprt *xprt = req->rq_xprt;
1678
1679         if (test_bit(RPC_TASK_NEED_XMIT, &task->tk_runstate) ||
1680             test_bit(RPC_TASK_NEED_RECV, &task->tk_runstate) ||
1681             xprt_is_pinned_rqst(req)) {
1682                 spin_lock(&xprt->queue_lock);
1683                 xprt_request_dequeue_transmit_locked(task);
1684                 xprt_request_dequeue_receive_locked(task);
1685                 while (xprt_is_pinned_rqst(req)) {
1686                         set_bit(RPC_TASK_MSG_PIN_WAIT, &task->tk_runstate);
1687                         spin_unlock(&xprt->queue_lock);
1688                         xprt_wait_on_pinned_rqst(req);
1689                         spin_lock(&xprt->queue_lock);
1690                         clear_bit(RPC_TASK_MSG_PIN_WAIT, &task->tk_runstate);
1691                 }
1692                 spin_unlock(&xprt->queue_lock);
1693         }
1694 }
1695
1696 /**
1697  * xprt_release - release an RPC request slot
1698  * @task: task which is finished with the slot
1699  *
1700  */
1701 void xprt_release(struct rpc_task *task)
1702 {
1703         struct rpc_xprt *xprt;
1704         struct rpc_rqst *req = task->tk_rqstp;
1705
1706         if (req == NULL) {
1707                 if (task->tk_client) {
1708                         xprt = task->tk_xprt;
1709                         xprt_release_write(xprt, task);
1710                 }
1711                 return;
1712         }
1713
1714         xprt = req->rq_xprt;
1715         if (task->tk_ops->rpc_count_stats != NULL)
1716                 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
1717         else if (task->tk_client)
1718                 rpc_count_iostats(task, task->tk_client->cl_metrics);
1719         xprt_request_dequeue_all(task, req);
1720         spin_lock_bh(&xprt->transport_lock);
1721         xprt->ops->release_xprt(xprt, task);
1722         if (xprt->ops->release_request)
1723                 xprt->ops->release_request(task);
1724         xprt->last_used = jiffies;
1725         xprt_schedule_autodisconnect(xprt);
1726         spin_unlock_bh(&xprt->transport_lock);
1727         if (req->rq_buffer)
1728                 xprt->ops->buf_free(task);
1729         xprt_inject_disconnect(xprt);
1730         if (req->rq_cred != NULL)
1731                 put_rpccred(req->rq_cred);
1732         task->tk_rqstp = NULL;
1733         if (req->rq_release_snd_buf)
1734                 req->rq_release_snd_buf(req);
1735
1736         dprintk("RPC: %5u release request %p\n", task->tk_pid, req);
1737         if (likely(!bc_prealloc(req)))
1738                 xprt->ops->free_slot(xprt, req);
1739         else
1740                 xprt_free_bc_request(req);
1741 }
1742
1743 #ifdef CONFIG_SUNRPC_BACKCHANNEL
1744 void
1745 xprt_init_bc_request(struct rpc_rqst *req, struct rpc_task *task)
1746 {
1747         struct xdr_buf *xbufp = &req->rq_snd_buf;
1748
1749         task->tk_rqstp = req;
1750         req->rq_task = task;
1751         xprt_init_connect_cookie(req, req->rq_xprt);
1752         /*
1753          * Set up the xdr_buf length.
1754          * This also indicates that the buffer is XDR encoded already.
1755          */
1756         xbufp->len = xbufp->head[0].iov_len + xbufp->page_len +
1757                 xbufp->tail[0].iov_len;
1758         req->rq_bytes_sent = 0;
1759 }
1760 #endif
1761
1762 static void xprt_init(struct rpc_xprt *xprt, struct net *net)
1763 {
1764         kref_init(&xprt->kref);
1765
1766         spin_lock_init(&xprt->transport_lock);
1767         spin_lock_init(&xprt->reserve_lock);
1768         spin_lock_init(&xprt->queue_lock);
1769
1770         INIT_LIST_HEAD(&xprt->free);
1771         xprt->recv_queue = RB_ROOT;
1772         INIT_LIST_HEAD(&xprt->xmit_queue);
1773 #if defined(CONFIG_SUNRPC_BACKCHANNEL)
1774         spin_lock_init(&xprt->bc_pa_lock);
1775         INIT_LIST_HEAD(&xprt->bc_pa_list);
1776 #endif /* CONFIG_SUNRPC_BACKCHANNEL */
1777         INIT_LIST_HEAD(&xprt->xprt_switch);
1778
1779         xprt->last_used = jiffies;
1780         xprt->cwnd = RPC_INITCWND;
1781         xprt->bind_index = 0;
1782
1783         rpc_init_wait_queue(&xprt->binding, "xprt_binding");
1784         rpc_init_wait_queue(&xprt->pending, "xprt_pending");
1785         rpc_init_wait_queue(&xprt->sending, "xprt_sending");
1786         rpc_init_priority_wait_queue(&xprt->backlog, "xprt_backlog");
1787
1788         xprt_init_xid(xprt);
1789
1790         xprt->xprt_net = get_net(net);
1791 }
1792
1793 /**
1794  * xprt_create_transport - create an RPC transport
1795  * @args: rpc transport creation arguments
1796  *
1797  */
1798 struct rpc_xprt *xprt_create_transport(struct xprt_create *args)
1799 {
1800         struct rpc_xprt *xprt;
1801         struct xprt_class *t;
1802
1803         spin_lock(&xprt_list_lock);
1804         list_for_each_entry(t, &xprt_list, list) {
1805                 if (t->ident == args->ident) {
1806                         spin_unlock(&xprt_list_lock);
1807                         goto found;
1808                 }
1809         }
1810         spin_unlock(&xprt_list_lock);
1811         dprintk("RPC: transport (%d) not supported\n", args->ident);
1812         return ERR_PTR(-EIO);
1813
1814 found:
1815         xprt = t->setup(args);
1816         if (IS_ERR(xprt)) {
1817                 dprintk("RPC:       xprt_create_transport: failed, %ld\n",
1818                                 -PTR_ERR(xprt));
1819                 goto out;
1820         }
1821         if (args->flags & XPRT_CREATE_NO_IDLE_TIMEOUT)
1822                 xprt->idle_timeout = 0;
1823         INIT_WORK(&xprt->task_cleanup, xprt_autoclose);
1824         if (xprt_has_timer(xprt))
1825                 timer_setup(&xprt->timer, xprt_init_autodisconnect, 0);
1826         else
1827                 timer_setup(&xprt->timer, NULL, 0);
1828
1829         if (strlen(args->servername) > RPC_MAXNETNAMELEN) {
1830                 xprt_destroy(xprt);
1831                 return ERR_PTR(-EINVAL);
1832         }
1833         xprt->servername = kstrdup(args->servername, GFP_KERNEL);
1834         if (xprt->servername == NULL) {
1835                 xprt_destroy(xprt);
1836                 return ERR_PTR(-ENOMEM);
1837         }
1838
1839         rpc_xprt_debugfs_register(xprt);
1840
1841         dprintk("RPC:       created transport %p with %u slots\n", xprt,
1842                         xprt->max_reqs);
1843 out:
1844         return xprt;
1845 }
1846
1847 static void xprt_destroy_cb(struct work_struct *work)
1848 {
1849         struct rpc_xprt *xprt =
1850                 container_of(work, struct rpc_xprt, task_cleanup);
1851
1852         rpc_xprt_debugfs_unregister(xprt);
1853         rpc_destroy_wait_queue(&xprt->binding);
1854         rpc_destroy_wait_queue(&xprt->pending);
1855         rpc_destroy_wait_queue(&xprt->sending);
1856         rpc_destroy_wait_queue(&xprt->backlog);
1857         kfree(xprt->servername);
1858         /*
1859          * Tear down transport state and free the rpc_xprt
1860          */
1861         xprt->ops->destroy(xprt);
1862 }
1863
1864 /**
1865  * xprt_destroy - destroy an RPC transport, killing off all requests.
1866  * @xprt: transport to destroy
1867  *
1868  */
1869 static void xprt_destroy(struct rpc_xprt *xprt)
1870 {
1871         dprintk("RPC:       destroying transport %p\n", xprt);
1872
1873         /*
1874          * Exclude transport connect/disconnect handlers and autoclose
1875          */
1876         wait_on_bit_lock(&xprt->state, XPRT_LOCKED, TASK_UNINTERRUPTIBLE);
1877
1878         del_timer_sync(&xprt->timer);
1879
1880         /*
1881          * Destroy sockets etc from the system workqueue so they can
1882          * safely flush receive work running on rpciod.
1883          */
1884         INIT_WORK(&xprt->task_cleanup, xprt_destroy_cb);
1885         schedule_work(&xprt->task_cleanup);
1886 }
1887
1888 static void xprt_destroy_kref(struct kref *kref)
1889 {
1890         xprt_destroy(container_of(kref, struct rpc_xprt, kref));
1891 }
1892
1893 /**
1894  * xprt_get - return a reference to an RPC transport.
1895  * @xprt: pointer to the transport
1896  *
1897  */
1898 struct rpc_xprt *xprt_get(struct rpc_xprt *xprt)
1899 {
1900         if (xprt != NULL && kref_get_unless_zero(&xprt->kref))
1901                 return xprt;
1902         return NULL;
1903 }
1904 EXPORT_SYMBOL_GPL(xprt_get);
1905
1906 /**
1907  * xprt_put - release a reference to an RPC transport.
1908  * @xprt: pointer to the transport
1909  *
1910  */
1911 void xprt_put(struct rpc_xprt *xprt)
1912 {
1913         if (xprt != NULL)
1914                 kref_put(&xprt->kref, xprt_destroy_kref);
1915 }
1916 EXPORT_SYMBOL_GPL(xprt_put);